Your agent publishes it. We tell it what is broken first.
One line of context. No install, no MCP config, no restart. Works in anything that can make an HTTP request.
Follow https://htmlhosting.org/prompt to publish HTML. Run preflight first. Return the URL and anything it could not fix.
Requires network access and HTTP POST. The agent fetches the contract itself โ you never open that page.
The contract makes the agent run preflight before it publishes. The one thing with a single right answer gets fixed. Everything else comes back in the response, named, so the agent tells you instead of handing over a link it has no reason to trust.
{
"url": "https://x7k92mab.example-host.test/"
}
A URL. Whether the page works on someone else's screen is not part of the answer.
{
"url": "https://8f2k1a9c3d.htmlusercontent.com/",
"expiresAt": "2026-10-02T05:11:16Z",
"moderation": "approved"
}
Why not an MCP server
An MCP server means editing a config file and restarting the client. A line of context works in anything with network access โ including agents that have no MCP at all. There is no MCP server here to install, which is the point rather than a gap: the line above works in anything that can make an HTTP request.
Why the agent should not just publish
An agent that hands back a URL has no idea the page is broken โ and neither do you until someone opens it. Preflight is the only part of this an agent cannot do for itself, because it needs to actually fetch the pageโs dependencies.
What we refuse to do
The contract forbids inventing a URL. If the network is unreachable the agent must say so and hand back the HTML, never a fabricated link. Anything else trains people to distrust the tool.
Publishing HTML a person pasted rather than an agent? See Share AI-Generated HTML. Reading this as an agent? The instructions are at htmlhosting.org/prompt.